Replies: 11 comments 40 replies
-
|
dimension_generalizer → propagate_shapes 举个例子： 这种风格导致其他人没法和它协作，尤其是我们后续还需要建立维度泛化的同源关系（即都源自于同一样本，比如samples/timm/resnet18）。这显然也会让与propagate_shapes的协作进行不下去。 解决办法重构dimension_generalizer的逻辑，让给定 这样后面所有的流程只需要把$output_dir/0/，$output_dir/1/等看成不同的workspace就行。同时同源关系也特别清楚。 |
Beta Was this translation helpful? Give feedback.
-
|
好滴，然后就是test/shape_propagator_test.sh的产物如图，哪里会用到呢？ |
Beta Was this translation helpful? Give feedback.
-
|
子图形状推导要用。看test/subgraph_input_shapes_naive_rewriter_test.sh。 |
Beta Was this translation helpful? Give feedback.
-
|
子图维度泛化后的输入形状信息不是推导出来的。而是“被赋予”的。现有整图的形状推导信息，然后按整图子图关系（input_producer_indexes）一块一块的“切”出来。 |
Beta Was this translation helpful? Give feedback.
-
对的，找到了他的引用了
|
Beta Was this translation helpful? Give feedback.
-
|
这里还有些数量关系需要梳理清楚。
9这个数字的来源是维度泛化会将一份旧样本产生9份新的样本。 |
Beta Was this translation helpful? Give feedback.
-
|
好的 |
Beta Was this translation helpful? Give feedback.
-
|
就算维度泛化产生了100倍的更多输入组合，我们的子图分解流程和整图子图关系（subgraph_sources.json和subgraph_input_producer_index等信息）还是不会变多的。
|
Beta Was this translation helpful? Give feedback.
-
|
好滴 |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
修改以后运行示例，现在就是output_dir/(0-8)/samples...... |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
修改graph_net/dimension_generalizer.py的_get_to_model_path方法，使用数字索引代替复杂的符号维度命名 |
Beta Was this translation helpful? Give feedback.
-
test/workspace_group_ranges_from_subgraph_sources 少了这个 |
Beta Was this translation helpful? Give feedback.
-
python
function group_subgraph_sources() {
echo ">>> [6] Group ranges from subgraph sources."
python3 -m graph_net.apply_sample_pass \
--model-path-list ${range_decomposed_subgraph_list} \
--sample-pass-file-path "$GRAPH_NET_ROOT/graph_net/sample_pass/group_ranges_from_subgraph_sources.py" \
--sample-pass-class-name GroupRangesFromSubgraphSources \
--sample-pass-config $(base64 -w 0 <<EOF
{
"subgraph_model_path_prefix": "${RANGE_DECOMPOSE_OUTPUT_DIR}",
"output_dir": "/tmp/workspace_group_ranges_from_subgraph_sources"
}
EOF
)
}
function analyze_subgraph_dependences(){
python3 -m graph_net.apply_sample_pass \
--model-path-list ${new_model_list} \
--sample-pass-file-path "$GRAPH_NET_ROOT/graph_net/torch/sample_pass/subgraph_input_producer_indexes_generator.py" \
--sample-pass-class-name SubgraphInputProducerIndexesGenerator \
--sample-pass-config $(base64 -w 0 <<EOF
{
"model_path_prefix": "${new_model_prefix}",
"subgraph_ranges_json_root": "/tmp/workspace_group_ranges_from_subgraph_sources",
"subgraph_ranges_json_file_name": "grouped_ranges_from_subgraph_sources.json",
"subgraph_ranges_json_key": "grouped_ranges_from_subgraph_sources",
"output_dir": "/tmp/workspace_subgraph_input_producer_indexes_generator"
}
EOF
)
} |
Beta Was this translation helpful? Give feedback.
-
|
你可以补充一下。这应该用于辅助子图输入形状泛化 |
Beta Was this translation helpful? Give feedback.
-
|
看函数的输入，analyze_subgraph_dependences是需要group_subgraph_sources的输出的 |
Beta Was this translation helpful? Give feedback.
-
|
main() { } main 所以顺序应该是
|
Beta Was this translation helpful? Give feedback.
-
|
整图泛化只是为了形状推导 不会基于泛化后的整图再做子图分解的 |
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
目前我的测试sh的流程 |
Beta Was this translation helpful? Give feedback.
-
|
子图输入生产者下标生成用到了整图，这里线没画出来 |
Beta Was this translation helpful? Give feedback.
-
|
这里的逻辑大体没错 |
Beta Was this translation helpful? Give feedback.
-
测试完毕
|
Beta Was this translation helpful? Give feedback.
-
|
graph_net/sample_pass/group_ranges_from_subgraph_sources.py 这个代码的逻辑有问题，我修改了 |
Beta Was this translation helpful? Give feedback.
-
|
graph_net/dimension_generalizer.py 这个代码也修改了 |
Beta Was this translation helpful? Give feedback.
-
文件均正常生成，但是未考量过转化率问题
|
Beta Was this translation helpful? Give feedback.
-
|
记得跑出来转化率 |
Beta Was this translation helpful? Give feedback.
-
|
后续会测试转化率 |
Beta Was this translation helpful? Give feedback.
-
add3c59663a1736364048c3943e5b7094c138df143dc670497c5aa8640934f48 |
Beta Was this translation helpful? Give feedback.
-
|
！突然好想觉得没错，我重新跑一下试试 |
Beta Was this translation helpful? Give feedback.
-
|
不行，运行失败的，疑问：为什么hash相同的要把子图范围合并？ 意义是什么？ |
Beta Was this translation helpful? Give feedback.
-
|
group_ranges_from_subgraph_sources.py 就是这个问题，需要排序，否则报错 |
Beta Was this translation helpful? Give feedback.
-
hash相同显然意味着同一个子图在多个地方出现过。所以需要合并。 subgraph_sources的信息是放在子图上的。它关注自己出现在哪些整图的哪些区域。 而group_ranges_from_subgraph_sources.py的作用是把子图的这些信息聚合到整图，每个整图都会维护一份信息表明哪些子图区域被抽取成了子图。 |
Beta Was this translation helpful? Give feedback.
-
|
现在结构就很清晰了 |
Beta Was this translation helpful? Give feedback.
-
|
我感觉我们的整图子图关系数据整得有点拧巴，这个首要责任在我。我在想是不是存成一个关系型数据的表是不是就很清楚了。他的schema就是full_graph_relative_model_path:str, subgraph_relative_model_path:str, subgraph_start:int, subgraph_end: int |
Beta Was this translation helpful? Give feedback.
-
|
我觉得要注重input和output以及他们各自的意义讲清楚，就像上文提到的为什么需要合并，我一开始的理解是节省空间，反正hash一样，到时候直接用这个model.py就可以了，但是其实意义是表示一个这样类型的子图在多段中出现过，即“它关注自己出现在哪些整图的哪些区域” |
Beta Was this translation helpful? Give feedback.
-
|
重新拉取最新代码跑失败了 原始模型： |
Beta Was this translation helpful? Give feedback.
-
|
更新了#589 |
Beta Was this translation helpful? Give feedback.
-
|
这里原始模型和泛化后模型 |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
这种有的是不需要转换pass、有的是pass变换后模型试跑失败了吧。可以验证下
这两种情况占比大吗？device有解决方案，但是不排除引起其他问题的可能，所以看下占比多大？ |
Beta Was this translation helpful? Give feedback.
-
试了下
small100里面的模型维度泛化后有108个样本出现了device 不一致问题： |
Beta Was this translation helpful? Give feedback.
-
|
1、对于最原始的GraphNet/graph_net/tools/generate_subgraph_dataset.sh代码，无任何修改,运行失败 如果cumsum_num_kernels_generator从子图上分析，分析出切分点，那我的fusible_subgraph_ranges_generator就应该根据切分点划分出整图的range范围，也就是samples/ultralytics/yolov3-tinyu下有一个切分文件而不是samples/timm/mobilenetv4_conv_aa_large.e230_r384_in12k/_decomposed/mobilenetv4_conv_aa_large.e230_r384_in12k_start108_end156_2 下有一个文件 修复PR：#617 2、修复以后可以在原版GraphNet/graph_net/tools/generate_subgraph_dataset.sh运行 ===================================================== ==================================================== |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
目前维度泛化相关的功能砖块有：
grpah_net.json里的dim_gen_pass字段tools/init_input_tensor_constraints.shtools/statisticize_in_tensor_symbolic_shapes.shtools/apply_dim_gen_passes.shtest/shape_propagator_test.shtest/subgraph_input_producer_indexes_generator_test.shtest/subgraph_input_shapes_naive_rewriter_test.shBeta Was this translation helpful? Give feedback.
All reactions